Dual-port 200Gb/s connectivity with aggregate 400Gb/s bandwidth, PCIe Gen 4.0 x16 interface, and low-profile form factor

The NVIDIA MCX653106A-HDAT ConnectX-6 InfiniBand smart adapter delivers dual-port 200Gb/s connectivity with 400Gb/s aggregate bandwidth. Featuring QSFP56 ports and PCIe Gen 4.0 x16 host interface, this low-profile card provides ultra-low latency, high message rate (215 million messages/sec), and advanced hardware offloads for HPC, AI, and enterprise data center workloads.

Built on the NVIDIA Quantum InfiniBand platform, ConnectX-6 integrates In-Network Computing engines, RDMA with end-to-end flow control, and block-level XTS-AES encryption. The adapter supports both InfiniBand and Ethernet protocols for flexible deployment across compute and storage infrastructures.

• Dual-port 200Gb/s (400Gb/s aggregate bandwidth)
• Up to 215 million messages per second
• Ultra-low latency performance
• Block-level XTS-AES encryption (256/512-bit key)
• FIPS compliant security
• PCIe Gen 4.0 and Gen 3.0 compatible
• In-Network Compute acceleration engines
• RDMA, send/receive semantics
• Hardware-based congestion control
• NVMe-of target and initiator offloads
• RoHS compliant, ODCC compatible
• Supports both 50G SerDes (PAM4) and 25G SerDes (NRZ)

ConnectX-6 offloads collective operations and MPI tag matching to the network, freeing CPU cycles for computation. Combined with NVIDIA PeerDirect RDMA (GPUDirect), it enables direct GPU-to-GPU communication for AI and machine learning workloads.

Hardware offload of IEEE AES-XTS encryption/decryption saves latency and CPU utilization, with dedicated encryption keys for multi-tenant protection.

Target and initiator offloads optimize NVMe over Fabrics, enhancing CPU utilization and scalability for storage environments.

Protocol Support: InfiniBand (200/100/50/40/25/10Gb/s) and Ethernet (200/100/50/40/25/10/1Gb/s)

• High-performance computing (HPC) clusters
• AI and machine learning training infrastructure
• Hyperscale cloud and enterprise data centers
• NVMe-of storage systems
• Virtualized networks including NFV
• x86, Power, Arm, GPU, and FPGA-based platforms
• Host chaining for economical rack design

Two 200Gb/s ports deliver massive bandwidth for demanding workloads.

XTS-AES block-level offload saves CPU cycles and ensures data-in-motion security.

Offloads MPI, tag matching, and collective operations to the network.

Target and initiator offloads enhance storage performance and scalability.

Compatible with major OS, virtualization platforms, and HPC middleware.

• Switches: NVIDIA Quantum-2, Quantum, Spectrum
• Cables: QSFP56 DAC / AOC / optics
• OS: Linux (RHEL, SLES, Ubuntu), Windows, VMware, FreeBSD
• Virtualization: SR-IOV, ASAP2, OVS
• Storage: NVMe-of, SRP, iSER, SMB Direct, NFS RDMA

• Confirm PCIe slot: x16 physical required
• Verify protocol: InfiniBand vs Ethernet
• Choose appropriate QSFP56 cables/optics
• Check bracket type: tall vs short
• Confirm driver/OS compatibility
• Verify port count required (dual-port vs single-port)

• Observe ESD handling precautions during installation
• Verify PCIe slot compatibility (x16 physical)
• Use NVIDIA-qualified QSFP56 optics and cables
• Ensure adequate airflow around the adapter
• Install appropriate drivers (OFED/WinOF-2) before use